text2vid img2vid H3 Minimax - Fully uncensored local AI video generator released today

plug your load image node to "Scale image to total pixels", then plug "get image size" to "image to video h3" node (default comfy mini max i2v workflow)
 
Does anyone know if H3 Minimax support img2img? First time using it and ComfyUI so any help would be great.
 
What prompt do you guys use for dildos, eg dildo handjob. The ones it generates are deformed and weird
 
H3 is more of a video model, for img2img Flux2.Klein or Krea2 are probably a better fit. If you're familiar with civitai, the easiest thing is to look at the popular workflows for the model and find one you like.
 
If you have questions I would highly recommend checking wildminders awesome minimax h3 page before asking here, great resource.

Hope this is okay to post as it has helped me quite a bit.
 
My eyes glazed over when I just looked his stuff up

Very overwhelming for the layman
 
@theAIsearch on youtube usually has great info for new models and usually a deeper dive into the bigger releases along with basic comfyui setup.
 
Minimax said they will be releasing an I2I model, but it's not out yet. Also, I've seen some workflows that use I2V to rip frames to serve as a pseudo-I2I, but when I tried it the quality was not good.
 
As a layman/very normie user myself will note everything I used to get myself setup this morning relatively quickly with never using ComfyUI or these models before. Able to consistently generate prompts for 5 second one shot I2V's that take about 5-8 minutes each with results I'm already happy with, have not worked with 5+ second video generation or a collection of LORA but feel this is a pretty good base to append this setup further for that usage. I have a upper mid tier GPU, and 32 GB RAM.

YT Videos:
- ComfyUI Course - Learn ComfyUI From Scratch | Full 5 Hour Course (Ep01) (specifically Chapter 3 and beginning of Chapter 4)
Note: This video is great because this an a portable install of ComfyUI that will come with QOL options like a launcher, and addons that will be prereq's for the second video. Essentially will be a clean install of ComfyUI that doesn't need to go on your C drive (assuming you have another drive you can use). When you install using EasyInstall you WILL NOT have to install ComfyUI from the website

- ComfyUI MiniMax H3: Best Video Generation Workflows (Ep29)
Note: particularly I am using the workflow from the 'H3 First Frame Workflow + More Settings' chapter

Other:

https://civitai.com/models/2834106/minimaxh3-auto-prompter-v73
This is a seperate ComfyUI workflow that is for the actual prompt generation. Found it from this same thread. The workflow will note the details/objects you need. This came from me not wanting to go through a separate LLM, and keep the prompt generation activity offline. For any single image I just pass the image, give a brief plain english description, and in about 20-30 second gives me a properly structured prompt for the above workflow.

Claude (Pro Sub tier): Used this as I encountered errors in running (and yes you will probably run into some), I did not have to specify the nature of what I was doing outside ComfyUI workflow troubleshooting, simply pasted errors and gave brief plan english context to files and such to assist. This bit is still a bit dependent on your comfortability with general AI usage. Also, am obviously not saying you need Claude, but an AI chatbot on the side goes a long way instead of having to dig through forums for very specific errors in what is already a niche subject.

"Technical" actions I did: Basic file explorer usage, edited some BAT files, ran PIP commands for Python stuff, and troubleshooting with AI (Claude as mentioned above). Everything else the workflows and videos themselves will outline what you need to do.

My workflow as is: I input an image to my prompt generation workflow, give a plain english description of what I want (about 3-5 sentences). Let it run and it gives me a structured prompt tailored to that exact photo. (This naturally means that every photo will have its own prompt. Depending on your use case you can use the structured prompts to perhaps create a generalized prompt if you'd like.) Paste that image and prompt into my actual H3 First Frame Workflow, config as needed, and boom I have my generated video. For LORA usage I plan to use Chapter 19 in the first video I mentioned to simply append my current workflow. Total setup time: About 2.5 hours (largely due to download times of the models).
 
eros10 works good, my only issue is that when I try to use video references it remakes the video but does not change the subject at all, idk if I'm doing something wrong or else, is it working for you guys?
 
that one is not the best --it has issues ((in my tests)) --your better off with a hybrid or try a different finetune
 
It's ok for anything involving sex and much better then the base model which is terrible without resource heavy video references
 
can someone explain how to get workflow ref2va? i got runpod with minimax uncensored , here re2va is 2 reference images not reference video . i tried to add video node manualy , but still cant generate anything properly , created promt with grok to swap person on video , it works with 1 photo then it dont work with other 20 , if someone can explain what am i doing wrong thank you
 
Have anyone found any good loras for soft pale skin, HMbreasts give more udders if you know what I mean.
 
are there any simple workflows which are speed focused without compromising the "originality" of reference image by too much that work on 8gb cards .Also What is the minimum time one can expect from such workflows .
 
Speed varies based on video resolution and video length, so smaller (0.30-0.40 megapixels) and shorter (3-4 seconds) will get you faster results. Larger resolution will keep from losing the "originality", so either choose a close-up reference image or increase resolution but sacrifice length.
 
Using Ref2VA to replace a character in a video.

Source video:


Final result:


Image of Chun-Li for
Source video above for
 
You must log in or register to reply here.
Back
Top